deepSRAMP

SRAMP is a popular mammalian m6A site predictor we previously developed (Nucleic Acids Res 2016). SRAMP has been totally cited by more than 570 papers (Google Scholar, 4-16, 2024) and represents the mostly used algorithm in this field. A large number of m6A sites were identified by the helps of SRAMP.

After ~8 years after its development, Now we released deepSRAMP (www.cuilab.cn/deepsramp) , which is designed based on a combined framework of transformer neural network and recurrent neural network by fusing the sequence and genomic position features. The results showed that SRAMP2 greatly outperforms its predecessor SRAMP with 15.0% increase of AUROC and 30.9% increase of AUPRC, and greatly outperforms other state-of-the-art m6A predictors (WHISTLE and DeepPromise) with average 16.1% and 18.3% increase of AUC and 43.9% and 46.4% increase of AUPRC, respectively.

Requirements

• torch
• pandas
• scikit-learn
• joblib
• tqdm
• matplotlib
• seaborn
• shap

Installation

1. Install conda and create a virtual enviroument named sramp with python installed.

conda create -y -n sramp python 
conda activate sramp

2. Clone this repo.

git clone https://github.com/zhfanrui/deepSRAMP.git

3. Install this package and dependencies.

cd deepSRAMP
pip install .

Usage

Inference

1. Download BLAST and Muscle.

sh setup_inference.sh

2. Prepare your query fasta and run

deepsramp predict \
--fasta /path/to/fasta.fa \
--db /path/to/database/hg38_mature \
--blast /path/to/blast/bin \
--model /path/to/model/full_400_mature.model \
--out ./result.csv

Train

1. Download GTF and FASTA files for training;

sh download.sh

2. See tutorials for training.

<!-- 2. Install `deepSRAMP` through ```sh pip install deepsramp ``` --> <!-- , especially for `data` and `model` folder; -->

Pretrained Models

Model names follow the rule of {mode}_{half_window_size}_{target}_{extra}.model.

| Model Name | Model Mode | Half Window Size | Window Size | Target | Extra | | :---: | :---: | :---: | :---: | :---: | :---: | | Main Models | | full_400_ythdf.model | Full Model | 400 | 801 | YTHDF1/2 | - | | full_400_mature.model | Full Model | 400 | 801 | Mature transcripts | - | | full_400_full.model | Full Model | 400 | 801 | Full transcripts | - | | Other Model Examples | | seqonly_400_mature.model | Sequence Features Only Model | 400 | 801 | Mature transcripts | - | | genomeonly_400_mature.model | Genome Features Only Model | 400 | 801 | Mature transcripts | - | | full_100_mature.model | Full Model | 100 | 201 | Mature transcripts | - | | full_600_mature.model | Full Model | 600 | 1201 | Mature transcripts | - | | full_400_mature_for_time.model | Full Model | 400 | 801 | Mature transcripts | For training time estimation | | Single Model Examples | | full_400_a549_single.model | Full Model | 400 | 801 | A549 | deepSRAMP$_{single}$ | | full_400_a549_dp.model | - | 500 | 1001 | A549 | DeepPromise | | full_400_a549_single_random.model | Full Model | 500 | 1001 | A549 | Selecting random transcript for training instead of max length trascript |

Tutorials

The reproduction of figures in the paper can be found in ipynb files.

• Figure 1
• Figure 2
• Figure 3

Citation

[paper]

Performance

Cross Validation

Test on YTHDF1/2

Test on m6Aatlas

<!-- rsync -av --exclude-from ../../deepsramp/exclude_file.txt ../../deepsramp/ . -->